Examples

Output Examples

Type	File
Input video	assets/onizuka_idle_motion.mp4
Animated WebP	example/output_animated.webp
GIF	output/output.gif
MatAnyone WebP 2 FPS / 300 px	output/matanyone_full_2fps_300.webp
MatAnyone WebP 5 FPS / 300 px	output/matanyone_full_5fps_300.webp
MatAnyone WebP 10 FPS / 300 px	output/matanyone_full_10fps_300.webp
MatAnyone GIF 10 FPS / 300 px	output/matanyone_full_10fps_300.gif
Comparison GIF	example/onizuka_walk_motion.gif
Comparison WebP	example/onizuka_walk_motion.webp
Transparent frames	output_frames_webp/

MatAnyone2 Tile WebUI Samples

The Tile workflow now documents both the UI and the generated split animations. The screenshots below were captured from the MatAnyone2 Tile tab, and the sample animated assets come from the documented 3x3 run under docs/public/media/matanyone2_tile/.

Resume from an existing Tile run

Preview grid after split export

Sample 3x3 tile outputs

Animated WebP tile	Animated GIF tile

Asset paths:

• /media/matanyone2_tile/webui-resume-en.png
• /media/matanyone2_tile/webui-preview-en.png
• /media/matanyone2_tile/tiles_3x3/tile_01_animated.webp
• /media/matanyone2_tile/tiles_3x3/tile_01_animated.gif

MatAnyone Recipes

Transparent WebP from a foreground + alpha pair

python main.py assets/MatAnyone --matanyone output/matanyone.webp

Compact preview: 5 FPS at 300 px

python main.py assets/MatAnyone --matanyone output/matanyone_5fps_300.webp --webp-fps 5 --size 300x300

Smaller preview: 2 FPS at 300 px

python main.py assets/MatAnyone --matanyone output/matanyone_2fps_300.webp --webp-fps 2 --size 300x300

Smoother preview: 10 FPS at 300 px

python main.py assets/MatAnyone --matanyone output/matanyone_10fps_300.webp --webp-fps 10 --size 300x300

Animated GIF at 10 FPS

python main.py assets/MatAnyone --matanyone output/matanyone_10fps_300.gif --animated gif --webp-fps 10 --size 300x300

Flatten to MP4 with a white background

python main.py assets/MatAnyone --matanyone output/matanyone.mp4 --bg-color white

Notes for MatAnyone inputs

• The alpha video controls which pixels stay visible: white stays, black becomes transparent.
• The exporter removes green matte contamination from semi-transparent edges before saving.
• 300x300 plus 5fps is usually a strong quality-to-size balance for previews.

Green Fringe Cleanup Comparison

This preview compares four MatAnyone cleanup strengths on assets/MatAnyone_cat3. soft keeps the fullest silhouette, while strong and trim suppress more visible green spill on whiskers and tails.

The residual map below highlights edge pixels where green is still stronger than red or blue after cleanup. It helps show the remaining hotspots instead of relying only on visual inspection.

• When saving multiple tuning candidates, include the profile or parameter token in the filename, for example MatAnyone_cat3_trim_sm0_gb4_rb60_as180_am120_md255.webp.
• In the filename tokens, sm=spill margin, gb=green bias, rb=red boost, as=alpha spill, am=alpha matte, and md=max drop.

Experiment: Fire Effect Clip

Test clip: assets/onizuka_fire_motion.mp4

Settings used:

python main.py assets/onizuka_fire_motion.mp4 output/model.webp --animated webp --webp-fps 8 --model <model>

Summary

• silueta produced the best overall balance on this clip.
• u2net kept the silhouette stable, but removed most of the fire aura.
• u2net_human_seg was not suitable for this stylized, effect-heavy sample.

Model Previews

Model	Preview	Runtime	Notes
isnet-general-use		114.44s	Preserves some effect detail, but leaves halo noise around the subject.
u2net		76.42s	Strong silhouette stability, but most of the fire aura disappears.
u2netp		30.11s	Fastest run, but quality drops on harder fire frames.
u2net_human_seg		69.97s	Loses most of the character on this stylized clip.
silueta		69.27s	Best balance of shape retention and cleanup in this comparison.

Visual Comparison Sheet

Representative frames sampled at 1.0s, 3.0s, 5.0s, 7.0s, and 9.0s. This makes it easier to spot where the aura survives and where the subject breaks apart.

Mask Comparison Sheet

The alpha-mask view is useful for checking subject coverage. u2net_human_seg collapses on this clip, while silueta, u2net, and isnet-general-use keep a far more complete silhouette.

Re-run the Experiment

The tracked experiment definition is stored under experiments/onizuka_fire_motion/.

• Script: experiments/onizuka_fire_motion/run_experiment.py
• Config: experiments/onizuka_fire_motion/experiment_config.json
• Notes: experiments/onizuka_fire_motion/README.md
• Output directory: output/model_experiments/onizuka_fire_motion/

Run it from the repository root:

python experiments/onizuka_fire_motion/run_experiment.py

To test an additional model later, add it to the models array in experiments/onizuka_fire_motion/experiment_config.json and run the same command again.

The script regenerates:

• <model>_anim.webp
• <model>_anim_frames/
• results.csv
• alpha_stats.csv
• comparison_sheet.png
• comparison_masks.png